Space group prediction of complex alloy systems by product-based neural networks

Abstract

Complex alloy systems exhibit some unique properties, many of which are attributed to ordering phenomena. At the atomic scale, this phenomenon refers to the occupancy probability of particles at the lattice sites. From the perspective of symmetry, it corresponds to different space groups. This study uses several machine learning algorithms to predict some common space groups of complex alloy systems. Under a large data set and a difficult classification task, the relevant models achieved excellent results on the test set. In the traditional support vector machine algorithm model, the prediction can reach the first-class level. Further, through the Product-based Neural Networks method under the wide and deep framework, the bottleneck of the traditional algorithm is broken through, and the prediction ability of the model is further improved. The average Area under curve value of the model can reach 99 %, and the prediction ability of multiple space groups has been improved. This study can not only provide more ideas for cross-scale modeling of complex systems, but the related models can also provide guidance for specific alloy design.